Joyus runs on one rule: a person directs the work, and every step is recorded, so you can check what was done instead of taking it on faith. Here is what that looks like in practice.
One session, start to finish
A person sets the task — review this pull request, draft this content, run this analysis. Joyus loads the skills that encode your standards: your coding conventions, your compliance rules, your voice, your review checklists, written down once and applied the same way every time. The work runs inside a governed boundary — policies and workflow gates you can read in the open repo — that decides what an agent may and may not touch, and every step it takes is recorded as a receipt. Before anything ships, a person reviews that record and signs off. Nothing reaches your systems on the model’s say-so alone.
The receipt is the point. When the work is done you do not just get an output; you get the trail of how it was produced — which skill ran, what it was allowed to do, what it changed, and who approved it. For regulated work, that trail is the difference between “we told the AI to behave” and “we can show you what it was allowed to do.”
Three freedoms that keep you in control
Choose your model. Joyus runs on a pluggable model layer. The same governed skill runs on Claude, Gemini, or OpenAI — or on a fully local, open-weight model when you would rather keep inference on your own infrastructure. You are not married to one vendor’s model or pricing. (Per-tenant bring-your-own-keys is on the roadmap; today the operator supplies the key.)
Inspect everything. The platform core is open under Apache-2.0, so you can read more than the marketing: the policies, the skills, the workflow gates, and the logs. What an agent is allowed to do is written where you can audit it, not buried in a prompt.
Leave whenever you want. Your specs, skills, and data are yours. The core is forkable today, and taking your work with you is a standard we hold — no one should be trapped because their know-how was encoded in Joyus.
What is real today, and what we are still building
We would rather tell you than dress it up. Real today: the open core, the pluggable model layer, the recorded receipts, and the governance that runs them. Still building: the closed learning loop — the intent is that every reviewed session makes the next one better — and per-tenant key management. The platform is alpha; the method is proven. The license means you are never trapped while we build.
Want this run, or built around your organization?
Joyus is the platform. The people who build it are Zivtech, a Claude consultancy that has spent 17 years building for healthcare, legal, education, and government. They run the discovery, encode your standards into skills, and can host and operate the whole thing for you. If you want help getting from “interesting” to “running in production,” see how Zivtech works with you on AI.